Method for Generating Customized Test Scenarios for an Emotional Driving-Based Personalized Driver Model

Abstract

The present invention relates to a method for generating customized test scenarios for a personalized driver model driven by emotions, including: based on the existing emotion model, selecting specific emotions that affect the behavior of driving a vehicle; pre-imposing specific emotion influences on the test driver, and then collecting the corresponding driving data of the test driver on a driving simulator; for the collected driving data, first performing imitation learning and then reinforcement learning Q-Learning to obtain generalized driving data of different styles; combining the driving data of different styles according to a specific ratio, and determining hyperparameters through a fireworks optimization algorithm, so as to generate an efficient test scenario for autonomous vehicles; determining the verification index of the efficient test scenario; selecting the decision-making method of the system under test, and verifying through the verification index. Compared with the prior art, the present invention can effectively improve the complexity and interpretability of the behavior of background vehicles in the test scenario, and improve the effectiveness and test efficiency of the test scenario.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving testing and evaluation, and in particular to a method for generating customized test scenarios for a personality driver model based on emotion-driven. Background Art

[0002] With the rapid development of autonomous driving technology, safety has become the primary basic issue to be solved in the field of autonomous driving. Among them, exploring test scenarios and enriching and improving test technologies are extremely important processes to improve the safety performance of autonomous driving. Considering that traditional real vehicle testing methods are difficult to achieve the diversity, coverage and typicality of test scenarios, and are inefficient, the simulation test scenario method is currently mainly used.

[0003] However, in existing simulation test scenarios, the behavior of background vehicles interacting with the test vehicle is often relatively fixed, and the form strategy followed is relatively simple, which does not conform to the complex and changeable driving environment in reality. In reality, driving behavior is often affected by the driver's driving habits, emotions during driving, and specific driving scenarios. Therefore, even vehicles that pass simulation tests still do not have good decision-making capabilities in actual driving scenarios; in addition, in current simulation test scenarios, the types of background vehicle behaviors are relatively small, which cannot accurately reflect the actual vehicle behavior, and it is difficult to explain the internal logic of different driving styles. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a method for generating customized test scenarios for a personalized driver model based on emotion-driven, which can improve the complexity and explainability of background vehicle behavior in the test scenario, and improve the effectiveness of the test scenario and the test efficiency.

[0005] The purpose of the present invention can be achieved by the following technical solution: A method for generating customized test scenarios for a personality driver model based on emotion-driven, comprising the following steps:

[0006] S1. Based on the existing emotion model, select specific emotions that affect driving behavior;

[0007] S2, applying specific emotional influences to the test driver in advance, and then collecting corresponding driving data of the test driver on the driving simulator;

[0008] S3: For the collected driving data, first perform imitation learning, then perform reinforcement learning Q-Learning to obtain generalized driving data of different styles;

[0009] S4. Combine driving data of different styles in a specific proportion and determine hyperparameters through the fireworks optimization algorithm to generate efficient test scenarios for autonomous vehicles.

[0010] S5. Determine the verification indicators for efficient test scenarios;

[0011] S6. Select a decision-making method for the system under test and verify it using the verification indicators determined in step S5.

[0012] Furthermore, the step S1 is specifically based on the SCCT (Situational Crisis Communication Theory) model, the Plutchik emotion wheel model and the OCC emotion model, from which six specific emotions that affect driving style are selected: anger, fear, doubt, happiness, contempt and tranquility, and based on the OCEAN five personality models, the six specific emotions are divided into three categories:

[0013] classifies anger, fear, and contempt as neurotic types;

[0014] Classify doubt and tranquility as types of responsibility;

[0015] Divide happiness into agreeableness types;

[0016] Among them, the impact of six specific emotions on driving behavior is as follows:

[0017] In a calm mood, the driver's driving style is stable, neither aggressive nor conservative, and is in a normal state;

[0018] When angry, drivers tend to adopt a more aggressive driving style, accelerate and decelerate more frequently, change lanes more frequently, and the probability of collision increases;

[0019] When drivers are afraid, they tend to drive more conservatively, accelerate and decelerate less frequently, change lanes less frequently, and the probability of collision is reduced.

[0020] Under the contempt emotion, due to the contempt for other people's driving skills or the contempt for the current road conditions, the driver tends to be slightly aggressive, with a slightly higher frequency of acceleration and deceleration and lane change than when calm, and the probability of collision is slightly increased;

[0021] When in doubtful mood, due to confusion and alertness about the road conditions and the behavior of other vehicles, the driver tends to be slightly conservative, with a slightly lower frequency of acceleration and deceleration and lane change than when in calm state, and a slightly lower probability of collision.

[0022] When drivers are in a happy mood, they tend to consider the feelings of other vehicle drivers, are more inclined to give way and reduce deceleration behavior that affects the vehicles behind, and the probability of a collision is lower.

[0023] Further, the specific emotions of anger, fear and doubt are selected from the SCCT model;

[0024] The specific emotions of contempt and tranquility are selected from the Plutchik Emotion Wheel Model;

[0025] The specific emotion of happiness is selected from the OCC Emotion Model.

[0026] Further, step S2 specifically includes the following steps:

[0027] S21. Initially induce emotions. By means of sensory stimulation, induce the test driver to reach six specific emotional states;

[0028] S22. Secondarily induce emotions. Guide the test driver to review the previously received sensory stimulation during driving;

[0029] S23. Hypothesize driving scenarios. Let the test driver test under different driving backgrounds, where the driving backgrounds include: normal commuting to and from work, going out for play, being cut off maliciously, being highly praised or belittled for the driver's driving skills, being informed that there are frequent car accidents on this section of the road;

[0030] S24. In a simulation environment, combine with a driving simulator to collect the driving data of the test driver after being induced by specific emotions at set time intervals;

[0031] S25. The test driver selects one emotion from the six specific emotions that he / she believes is induced. If the expected induced emotion is the same as the emotion selected by the test driver, the data collected in step S24 is considered valid data; otherwise, return to execute step S21.

[0032] Further, the means of sensory stimulation includes but is not limited to watching videos, listening to music, viewing pictures, reading text, playing games.

[0033] Further, step S3 specifically includes the following steps:

[0034] S31. Obtain the driving data of the collected test driver, where the driving data includes driving state and driving behavior;

[0035] S32. Combine the current driving state s i 、the current driving behavior a i 、the driving state s i ' of the next time step, and the current reward value r i , and construct a data set

[0036] S33. Conduct a batch sampling from the data set :

[0037]

[0038] s ← s′

[0039] where Q φ (s′ i , a′ i ) is the maximum reward for one time step, γ is the decay rate, α is the learning rate, [r(s i , a i ) + γ max a′ Q φ (s′ i , a′ i )] is the actual reward, Q φ (s i , a i ) is the estimated reward, s is the current state, s′ is the state reached after taking action a. At each time step, the reward for the current behavior and the future reward caused by the expected current behavior are calculated. At each time step, the current reward Q-table is updated. This loop continues until s reaches an extreme value.

[0040] Furthermore, step S4 specifically includes the following steps:

[0041] S41. Establish a road model containing six specific emotions with different proportions and store it in the form of an array;

[0042] S42. Randomly initialize N firework positions in the solution space;

[0043] S43. Obtain explosion sparks and Gaussian sparks;

[0044] S44. Determine whether the set test scenario efficiency condition is met. If it is met, end the loop; otherwise, return to step S43 for looping.

[0045] Furthermore, the specific process of step S43 is as follows:

[0046] First, determine that the number of fireworks generated by the explosion and the explosion radius are respectively:

[0047]

[0048] where T p is the number of sparks generated by the explosion of the p-th firework, m and d are constants, f(x p ) is the value of the individual fitness, Y max and Y min are respectively the maximum and minimum values of the current population fitness, and ε is a very small value;

[0049] After the fireworks explode, displacement operations and mutation operations need to be performed on the explosion sparks. The method of random displacement is used to update the dimensions of the sparks. The mutation operation is to expand the optimization space and increase the diversity of the population to avoid the algorithm falling into local optima. Gaussian mutation is used to generate mutant sparks:

[0050]

[0051] Among them, e follows a Gaussian distribution with a variance of 1 and a mean of 1;

[0052] After that, the strategy based on Manhattan distance in the standard fireworks algorithm is used to select the next generation of fireworks:

[0053]

[0054] Among them, d(xi,xj) is the Manhattan distance between two sparks.

[0055] Furthermore, the specific process of step S5 is as follows:

[0056] First, set the variables for constructing the verification metrics. The variables include the number of lane changes num_changelines, the number of accelerations num_acceleration, the number of decelerations num_deceleration, the absolute value of acceleration abs_acceleration, the absolute value of deceleration abs_deceleration, the interaction effect num_interaction, the expected time to collision num_timetocollision, and the number of collisions num_collision within one minute;

[0057] The range of each variable is [0,1], which is achieved by dividing the real-time value of the variable by the maximum value of the change amount in all driving data;

[0058] Allocate different weights to the above 8 variables and linearly add them. The obtained value represents the efficiency of generating the test scenario. The larger the value, the better the efficiency of generating the test scenario;

[0059] Among them, the allocated weight of the number of collisions num_collision is 0.3, and the remaining variables are set to equal weights according to the maximum entropy model, all of which are 0.1.

[0060] Furthermore, step S6 builds a simulation experiment platform based on PreScan software and Carsim software to detect each variable set in step S5 in real time and uses the decision tree algorithm for verification.

[0061] Compared with the prior art, based on multiple emotion models, the present invention selects specific emotions that affect driving behavior, and by imposing emotional influences on the test drivers and collecting corresponding driving data, driving behavior models under different emotions are obtained, thereby effectively enhancing the interpretability and predictability of the behavior of the vehicle ahead.

[0062] After obtaining the driving data, the present invention uses a method that combines imitation learning and machine learning. It can not only obtain behaviors superior to human driving data, making it easy to extract the characteristics of driving behaviors under different emotions, but also, compared with single methods such as supervised learning, the data processing of this method is more stable, less likely to have the situation of non-convergence of training data, and ensures that the behavior of the vehicle ahead is more in line with reality.

[0063] The present invention combines driving data of different styles in a specific proportion, determines the hyperparameters through the fireworks optimization algorithm, enriches the behavior strategies of the vehicle ahead, provides a driving environment closer to reality for testing, makes the test environment more complex. In addition, by setting variables, verification indicators for verifying efficiency are constructed, thereby realizing a method for verifying the efficiency of a test scenario, which is beneficial to improving the efficiency of the simulation test scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a schematic flowchart of the method of the present invention;

[0065] Figure 2 is a classification diagram of the emotion model in the embodiment;

[0066] Figure 3 is a flowchart of the emotion guidance method in the embodiment;

[0067] Figure 4 is a schematic diagram of the videos, games, and music selected for guiding emotions in the embodiment;

[0068] Figure 5 is a flowchart of combining imitation learning and Q-learning reinforcement learning in the present invention;

[0069] Figure 6 is a flowchart of executing the fireworks optimization algorithm to determine the hyperparameters in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0070] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] Embodiment

[0072] As Figure 1 shown, a method for generating a customized test scenario for a personalized driver model based on emotion driving includes the following steps:

[0073] Step S1: Based on the existing emotion model, select the emotions that affect the driving behavior of the vehicle;

[0074] Step S2: applying influences to the subject in advance in various ways, and then collecting corresponding driving data on a driving simulator;

[0075] Step S3: first perform imitation learning on the collected driving data, and then perform reinforcement learning Q-Learning;

[0076] Step S4: Combine driving data of different styles in a specific proportion, obtain appropriate hyperparameters through the fireworks optimization algorithm, and generate an efficient test scenario for autonomous driving vehicles;

[0077] Step S5: Determine verification indicators for efficient test scenarios;

[0078] Step S6: Select the decision mode of the system under test and verify it through the verification index determined in step S5.

[0079] Specifically, the induction process of the emotion model in step S1 is as follows: Figure 2 As shown in the figure, combining the crisis situation communication model SCCT, the OCC emotional model and the OCEAN five personality model, six specific emotions that affect driving style are selected: anger, fear, doubt, happiness, contempt and tranquility. The above six emotions are divided into three categories: anger, fear and contempt belong to neuroticism; happiness belongs to agreeableness, doubt and tranquility belong to responsibility.

[0080] The above classification is explained as follows: anger, fear and doubt are selected from the Situational Crisis Communication Theory (SCCT). According to the content of the SCCT, since the semantics of emotions such as surprise, worry, alertness, anxiety, disgust, and disdain in crisis situations overlap to a certain extent, and the number is relatively small in the comment data, they are included in the fear and anger categories. Therefore, the fear emotion in the present invention also includes vigilance, surprise, worry, etc., which make the driver adopt a conservative driving style, and the anger emotion also includes disgust, anxiety, etc., which make the driver adopt an aggressive driving style. Contempt and tranquility are selected from the Plutchik emotional wheel. Contempt includes contempt for other people's driving skills or contempt for current road conditions. Tranquility represents being in a normal driving environment, and driving behavior is neither radical nor conservative. Happiness is selected from the OCC emotional model, and driving behavior in a happy mood tends to be more friendly. Neuroticism, conscientiousness and agreeableness are selected from the OCEAN five personality models. Neuroticism: It is difficult to balance anxiety, hostility, depression, self-consciousness, impulsiveness, fragility and other emotional traits, that is, it is incapable of maintaining emotional stability; Conscientiousness: It shows competence, fairness, orderliness, diligence, achievement, self-discipline, caution, restraint and other characteristics; Agreeableness: It has trust, altruism, straightforwardness, compliance, modesty, empathy and other characteristics. According to its attributes, anger, fear and contempt belong to neuroticism; doubt and tranquility belong to conscientiousness; happiness belongs to agreeableness, so as to combine the six emotions in a more systematic way.

[0081] The influence of the six emotions on driving behavior is defined as follows: tranquility belongs to the sense of responsibility. Under this emotion, the driver's driving style is stable, neither aggressive nor conservative, and is in a normal state; anger belongs to the neurotic type. Under this emotion, the driver tends to have a more aggressive driving style, with a higher frequency of acceleration and deceleration, a higher frequency of lane changes, and a higher probability of collision; fear belongs to the neurotic type. Under this emotion, the driver tends to have a more conservative driving style, with a lower frequency of acceleration and deceleration, a lower frequency of lane changes, and a lower probability of collision; contempt belongs to the neurotic type. Under this emotion, due to contempt for other people's driving skills or Due to the indifference to the current road conditions, the driver tends to be slightly aggressive, and the frequency of acceleration and deceleration is slightly higher than when the vehicle is calm, and the frequency of lane changes is slightly higher, which slightly increases the probability of a collision; doubt belongs to the sense of responsibility. Under this emotion, due to a certain degree of confusion and vigilance about the road conditions and the behavior of other vehicles, the driver tends to be slightly conservative, and the frequency of acceleration and deceleration is slightly lower than when the vehicle is calm, and the frequency of lane changes is slightly lower, which slightly reduces the probability of a collision; happiness belongs to agreeableness. Under this emotion, the driver tends to consider the feelings of other vehicle drivers, and is more inclined to give way and reduce the deceleration behavior that affects the following vehicles, so the probability of a collision is lower than when the vehicle is calm.

[0082] In step S2, emotional influence can be imposed on the subject through means such as videos, music, pictures, texts, games, conversations, and simulated driving scenarios.

[0083] The emotional guidance process in step S2 is as Figure 3 shown:

[0084] Step S21: Initial guidance through the senses. Before the test, let the subject undergo emotional guidance listed in step S2, such as watching videos (violent videos, thriller videos, reasoning videos, landscape videos, racing videos, healing videos), listening to music (rock music, light music, etc.), viewing pictures (violent pictures, blurred pictures, car accident pictures, landscape pictures, racing pictures, healing pictures), reading texts (of the same type as pictures), playing games (violent games, thriller games, racing games, reasoning games, healing games, casual games), etc. The purpose of these multiple sensory stimulation methods is to guide the subject to reach the six emotional states selected in step S1. The videos, games, and music for emotional guidance in this embodiment are as Figure 4 shown;

[0085] Step S22: Secondary guidance. Guide the subject to briefly review the sensory stimuli received during the driving process;

[0086] Step S23: Hypothetical driving scenarios. Let the subject be tested under different driving backgrounds, including: normal commuting, going out for fun, being cut off maliciously, receiving high praise or criticism for the driver's driving skills, being informed that there are frequent car accidents on this section of the road, etc.;

[0087] Step S24: Combine with a driving simulator in a simulation environment to collect the driving behavior of the subject after receiving the above emotional guidance;

[0088] Step S25: The subject selects the one emotion from the six emotions listed in step S1 that he / she believes is the closest one generated by the guidance, to further verify the effectiveness of the emotional guidance. If the expected guided emotion does not match the emotion considered by the subject to be generated, repeat step S2. If it still does not match, the set of data is regarded as invalid.

[0089] Step S3 combines imitation learning and Q-learning reinforcement learning, and its process is as Figure 5 shown, including:

[0090] Step S31: Collect the demonstration data of the subject:

[0091] (s i ,a i )

[0092] Step S32: Initialize the data set and make it contain the demonstration data:

[0093] (s i ,a i )

[0094] Step S33: Collect the data set with the same strategy:

[0095] {(s i ,a i ,s i ′,r i )}, and add it to the data set ;

[0096] Step S34: Take a batch of samples from the data set ;

[0097]

[0098] s ← s′

[0099] where Q φ (s′ i ,a′ i ) represents the maximum return of one step, γ represents the decay rate, α represents the learning rate, [r(s i ,a i ) + γmax a′ Q φ (s i ′,a′ i )] represents the actual return, Q φ (s i ,a i ) represents the estimated return. For each step, calculate the current action return and the future return caused by the expected current action. For each step, update the current return Q table. Repeat this process until s reaches the extreme value;

[0100] Use a method that combines imitation learning and machine learning, aiming to combine the advantages of the two algorithms. The advantage of imitation learning is that the training process is very stable, and the disadvantage is that a large amount of artificial data needs to be provided and the data distribution is uneven; the advantage of reinforcement learning is that it can obtain data better than the existing human behavior, and the disadvantage is that it may not converge. If there is both artificial data and a reward function, the two can be combined. In this way, it is possible to perform behaviors better than human driving data, so it is easy to extract the characteristics of driving behaviors under different emotions. Compared with single methods such as supervised learning, the data processing of this method is also more stable and less likely to have the situation of non-convergence of training data.

[0101] The process of Step S4 is as Figure 6 shown, including:

[0102] Step S41: Establish a road model containing six driving emotions with different proportions and store it in the form of an array;

[0103] Step S42: Randomly initialize the positions of N fireworks in the solution space;

[0104] Step S43: Obtain explosion sparks and Gaussian sparks;

[0105] The number of fireworks generated by the explosion and the explosion radius are respectively:

[0106]

[0107] where, T p is the number of sparks generated by the explosion of the p-th firework, m and d are constants, f(x p ) is the value of the individual fitness, Y max and Y min are respectively the maximum and minimum values of the fitness of the current population, and ε is a very small value;

[0108] After the fireworks explode, displacement operations and mutation operations need to be performed on the explosion sparks. The method of random displacement is used to update the dimensions of the sparks. The mutation operation is to expand the optimization space and increase the diversity of the population to avoid the algorithm falling into local optimum. Gaussian mutation is used to generate mutant sparks:

[0109]

[0110] where, e follows a Gaussian distribution with a variance of 1 and a mean of 1;

[0111] After that, the strategy based on Manhattan distance in the standard fireworks algorithm is used to select the next generation of fireworks:

[0112]

[0113] where, d(xi,xj) is the Manhattan distance between two sparks.

[0114] Step S44: Determine whether the required scene efficiency condition is satisfied. If it is satisfied, end the loop; otherwise, continue the loop of step S43;

[0115] In step S5, set variables:

[0116] The number of lane changes num_changelines within one minute;

[0117] The number of accelerations num_acceleration;

[0118] The number of decelerations num_deceleration;

[0119] The absolute value of acceleration abs_acceleration;

[0120] Absolute value of deceleration, abs_deceleration;

[0121] Number of interactions, num_interaction;

[0122] Time to collision, num_timetocollision and number of collisions, num_collision.

[0123] Let the verification index be:

[0124] G = 0.1 * num_changelines + 0.1 * num_acceleration + 0.1 * num_deceleration + 0.1 * abs_acceleration + 0.1 * abs_deceleration + 0.1 * num_interaction + 0.1 * num_timetocollision + 0.3 * num_collision;

[0125] Among them, the number of interactions, num_interaction, represents the impact on the following vehicle due to the deceleration of the vehicle under test. Its value is the absolute value of the deceleration generated by the following vehicle due to the deceleration of the vehicle under test. The time to collision, num_timetocollision, represents the degree of danger of the current distance from the vehicle in front. The calculation method is as follows: Let a be the distance from the vehicle under test to the vehicle in front, b be the current speed of the vehicle under test, c = a / b, ttc = -exp(-ttc), that is, the change amount is the penalty term. When there is no vehicle in front, the penalty is zero, and the closer to the vehicle in front, the greater the penalty;

[0126] The range of each variable is [0, 1], which is achieved by dividing the real-time value of the variable by the maximum value of the variable that appears in all driving data; The G value represents the efficiency of the generated scenario. The larger the value, the better the efficiency of the generated test scenario.

[0127] Since a collision is more dangerous and important, the assigned weight is 0.3, and other variables are set to equal weights according to the maximum entropy model, all of which are 0.1;

[0128] Step S6 then builds a simulation experiment platform based on PreScan and Carsim software (where PreScan software is used for traffic scenario simulation and virtual sensor modeling, and CarSim software is used for vehicle dynamics modeling) to detect the variables set in step S5 in real time. The test system uses a decision tree algorithm for efficiency verification.

[0129] In summary, the present technical solution selects specific emotions that have a greater impact on driving behavior, resulting in significant differences in the driving styles of background vehicles under different emotions, and improving the complexity and interpretability of the behavior of background vehicles in the test scenario. By combining optimized learning of driving data, determining hyperparameters through the fireworks optimization algorithm, and setting variables to construct verification metrics, the efficiency of the test scenario and the test efficiency are further improved.

Claims

1. A method for generating customized test scenarios for a personality driver model based on emotion drive, characterized in that, The following steps are involved: S1. Based on the existing emotion model, select specific emotions that affect driving behavior; S2, applying specific emotional influences to the test driver in advance, and then collecting corresponding driving data of the test driver on the driving simulator; S3: For the collected driving data, first perform imitation learning, then perform reinforcement learning Q-Learning to obtain generalized driving data of different styles; S4. Combine driving data of different styles in a specific proportion and determine hyperparameters through the fireworks optimization algorithm to generate efficient test scenarios for autonomous vehicles. S5. Determine the verification indicators for efficient test scenarios; S6, select the decision-making method of the system under test, and verify it through the verification index determined in step S5; Step S1 is based on the SCCT model, the Plutchik emotion wheel model and the OCC emotion model, from which six specific emotions that affect driving style are selected: anger, fear, doubt, happiness, contempt and tranquility, and based on the OCEAN five personality models, the six specific emotions are divided into three categories: classifies anger, fear, and contempt as neurotic types; Classify doubt and tranquility as types of responsibility; Divide happiness into agreeableness types; Among them, the impact of six specific emotions on driving behavior is as follows: In a calm mood, the driver's driving style is stable, neither aggressive nor conservative, and is in a normal state; When angry, drivers tend to adopt a more aggressive driving style, accelerate and decelerate more frequently, change lanes more frequently, and the probability of collision increases; When drivers are afraid, they tend to drive more conservatively, accelerate and decelerate less frequently, change lanes less frequently, and the probability of collision is reduced. Under the contempt emotion, due to the contempt for other people's driving skills or the contempt for the current road conditions, the driver tends to be slightly aggressive, with a slightly higher frequency of acceleration and deceleration and lane change than when calm, and the probability of collision is slightly increased; When in doubtful mood, due to confusion and alertness about the road conditions and the behavior of other vehicles, the driver tends to be slightly conservative, with a slightly lower frequency of acceleration and deceleration and lane change than when in calm state, and a slightly lower probability of collision. When the driver is in a happy mood, he or she tends to consider the feelings of other drivers, is more inclined to give way and reduce the deceleration behavior that affects the following vehicles, and the probability of collision is lower; Step S3 specifically includes the following steps: S31, acquiring the collected driving data of the test driver, wherein the driving data includes driving status and driving behavior; S32. Combine the current driving state s i , the current driving behavior a i , the driving state s' at the next time step i and the current reward value r i to construct a data set S33. Take a batch of samples from the data set : s←s′ Among them, Q φ (s′ i , a′ i ) is the maximum return for a time step, γ is the decay rate, α is the learning rate, [r(s i , a i ) + γmax a′ Q φ (s′ i , a′ i )] is the actual return, Q φ (s i , a i ) is the estimated return, s is the current state, s′ is the state reached after taking the action a. For each time step, calculate the current action return and the future return caused by the expected current action. For each time step, update the current return Q-table. Repeat this process until s reaches an extreme value; Step S4 specifically includes the following steps: S41, establishing a road model including six specific emotions in different proportions, and storing it in an array form; S42, randomly initialize N firework positions in the solution space; S43, obtaining explosion sparks and Gaussian sparks; S44: Determine whether the set test scenario efficiency condition is met. If so, end the loop; otherwise, return to step S43 and loop.

2. The method for generating a customized test scenario of a personality driver model based on emotion driving according to claim 1, wherein The specific emotions of anger, fear and doubt are selected from the SCCT model; The specific emotions of contempt and tranquility are selected from the Plutchik emotion wheel model; The specific emotion of happiness is selected from the OCC emotion model.

3. A method for generating a customized test scenario for an emotion-driven personalized driver model according to claim 1, wherein, The specific steps of step S2 are as follows: S21. Initially guide the emotions. Adopt sensory stimulation methods to guide the test driver to reach six specific emotional states; S22. Secondarily guide the emotions. Guide the test driver to review the previously received sensory stimulation during driving; S23. Driving scenario assumption. Let the test driver test under different driving backgrounds, where the driving backgrounds include: normal commuting, going out for fun, being maliciously cut off, highly evaluating or demeaning the driver's driving skills, and informing that there are frequent car accidents on this section of the road; S24. Combine with a driving simulator in a simulation environment. Collect the driving data of the test driver after being guided by specific emotions at set time intervals; S25. The test driver selects one of the six specific emotions that he / she believes is induced. If the expected induced emotion is the same as the emotion selected by the test driver, the data collected in step S24 is considered valid data, otherwise return to execute step S21.

4. A method for generating a customized test scenario for a personality driver model based on emotion driving, as claimed in claim 3, wherein The sensory stimulation methods include but are not limited to watching videos, listening to music, viewing pictures, reading text, and playing games.

5. A method for generating a customized test scenario of a personality driver model based on emotion driving, characterized in that The specific process of step S43 is as follows: First, determine that the number of fireworks generated by the explosion and the explosion radius are respectively: Among them, T p is the number of sparks generated by the explosion of the p-th firework, m and d are constants, f(x p ) is the value of the individual fitness, Y max and Y min are respectively the maximum and minimum values of the fitness of the current population, and ε is a very small value; After the fireworks explode, displacement operations and mutation operations need to be performed on the explosion sparks. The method of random displacement is used to update the dimensions of the sparks. The mutation operation is to expand the optimization space and increase the diversity of the population to avoid the algorithm falling into local optimum. Gaussian mutation is used to generate mutant sparks: Among them, e follows a Gaussian distribution with a variance of 1 and a mean of 1; After that, the strategy based on Manhattan distance in the standard fireworks algorithm is used to select the next generation of fireworks: Among them, d(xi,xj) is the Manhattan distance between two sparks.

6. A method for generating a customized test scenario of a personality driver model based on emotion drive according to claim 1, characterized in that The specific process of step S5 is as follows: First, set the variables for constructing the verification indicators. The variables include the number of lane changes num_changelines, the number of accelerations num_acceleration, the number of decelerations num_deceleration, the absolute value of acceleration abs_acceleration, the absolute value of deceleration abs_deceleration, the interaction influence num_interaction, the expected time to collision num_timetocollision, and the number of collisions num_collision within one minute; The range of each variable is [0,1], which is achieved by dividing the real-time value of the variable by the maximum value of the change amount that appears in all driving data; Allocate different weights to the above 8 variables and linearly add them. The obtained value represents the efficiency of generating the test scenario. The larger the value, the better the efficiency of generating the test scenario; Among them, the allocated weight of the number of collisions num_collision is 0.3, and the remaining variables are set to equal weights according to the maximum entropy model, all of which are 0.

1.

7. A method for generating a customized test scenario of a personality driver model based on emotion driving, characterized in that, Step S6 builds a simulation experiment platform based on PreScan software and Carsim software to detect each variable set in step S5 in real time, and uses a decision tree algorithm for verification.

Images